Course details

8.3453

Multi-agent communication in a multi-task visual world

 SP
SS 2020 Dr. Elia Bruni
6h/wk
12 ECTS 		M.Sc modules:
CC-MP-SP - Study Project
CS-MP-SP - Study Project
KOGW-MPM-SP - Study project

Doctorate program 		Wed: 10-12
Thu: 10-12

A longstanding goal of AI is to develop artificial agents that can cooperate with humans to achieve goals using natural language communication. Recent studies have shown that artificial agents trained to solve a task using a discrete language-like channel succeed at learning to communicate in order to exchange information required to solve this task (e.g., [1-13]). However, the ‘languages’ they develop are not human-interpretable, and little is known about what sort of information agents are exchanging and how their syntactic and semantic spaces are organised. In this project, we will focus on training agents to communicate in a multi-task environment and analyse the properties of the languages that emerge. In particular, we are interested in what kind of representations agents learn and how to improve them, what kind of linguistic structure is present in the messages, how well agents generalise when confronted with out of distribution samples, how can we translate from 'neuralese' to human language, and many other research questions. Let's together define the most suitable research objective for you. Required skills - Independent and proactive attitude and an interest in artificial intelligence - Solid math background: calculus, linear algebra, probability and statistics - Good programming skills (PLUS: Pytorch or other DL libraries) - Knowledge and skill in developing and applying machine learning algorithms (MUST: deep learning. PLUS: reinforcement learning, evolutionary computation) and evaluating prediction performance (experience on some specific AI domain such as NLP, computer vision, etc) - We require many skills from you, but at the same time, it is of course perfectly normal if there are some topics or techniques that you still have to catch up on. Therefore don't worry, if we match, we will also offer you our help to efficiently fill possible gaps.